[codex] fix EasySession input buffer lifetime

[justinjoy]

Summary

Fixes a PyreWire FFI lifetime bug exposed by wirelog#863. EasySession.insert() and remove() were passing temporary Python-owned bytes and ctypes.c_int64[] buffers into the wirelog easy facade, then allowing those objects to be reclaimed before the next evaluating call. If Python bookkeeping or other allocations happened before step(), the C incremental path could consume stale row pointers and emit missing deltas.

This keeps easy-facade input buffers alive on the session until step() or snapshot() has consumed them, and clears the keepalive list afterward. The same keepalive treatment is applied to the variadic insert_sym/remove_sym path.

Validation

• .venv/bin/ruff check src tests passed
• WIRELOG_LIB=/home/joykim/git/semantic-reasoning/wirelog/builddir-861/libwirelog.so.1 PYTHONPATH=src .venv/bin/pytest -q -o addopts='' tests/test_easy_session.py tests/test_easy_step_snapshot.py tests/test_easy_deltas.py tests/test_easy_insert_sym.py tests/test_async.py passed: 48 passed
• Full WIRELOG_LIB=/home/joykim/git/semantic-reasoning/wirelog/builddir-861/libwirelog.so.1 PYTHONPATH=src .venv/bin/pytest -q reached 93.20% coverage but failed one existing metadata test: tests/test_changelog_format.py::test_topmost_released_section_matches_pyproject because CHANGELOG.md topmost released section is 0.41.0 while pyproject.toml is 0.41.99.

Notes

Raw ctypes calls into wirelog_easy_set_delta_cb/wirelog_easy_step with a manually retained row buffer produced the expected reachability deltas, confirming the C easy facade was not the failing layer in this reproduction.